BrainVision AI: A Privacy-First Solution for Brain Tumor Classification in the Browser

BrainVision AI is a browser-based brain tumor MRI image classification tool that uses client-side machine learning technology. Its core goal is to resolve the conflict between computational resource requirements and patient privacy protection in the field of medical AI. This tool enables real-time diagnostic assistance without server interaction—MRI images remain on local devices at all times. It features zero data transmission, instant response, offline availability, and cross-platform compatibility, providing a privacy-first intelligent service solution for clinical assistance, scientific research and teaching, and patient education.

Data Sensitivity Challenges

Medical images are among the most sensitive personal information. Traditional cloud-based AI diagnosis requires data uploads, which face issues of compliance, security, and patient trust. Additionally, they are subject to strict restrictions from regulations like the EU's GDPR and the U.S.'s HIPAA regarding cross-border transmission and third-party processing of medical data.

Edge Computing Solutions

The concept of edge computing pushes computation to where the data resides. BrainVision AI is a practical application of this concept in the medical imaging field, avoiding the migration of data to computing centers.

Technical Architecture

1. Client-side Inference Design: The model (trained with Keras → converted to ONNX format) runs in the browser via WebAssembly/JS engines, enabling zero data transmission, instant response, offline availability, and cross-platform compatibility.
2. Model Training and Export: Keras models are trained using public datasets, converted to ONNX format via tf2onnx, then compiled into WebAssembly modules to optimize performance.
3. UI Design: Responsive layout adapts to multiple devices, supporting multi-language and theme switching. The three-step operation process (upload → analyze → view results) is simple and user-friendly.

Functional Features

• Core Features: MRI image upload and analysis (supports JPG/PNG/BMP), real-time feedback, result export.
• Auxiliary Features: Training pipeline integration (supports training and fine-tuning with custom datasets), model version management.

Clinical Auxiliary Diagnosis

• Preliminary screening and triage: Primary institutions quickly identify suspected cases to guide referrals;
• Teaching and training: Case practice for medical students;
• Teleconsultation: Local preliminary analysis in network-restricted areas before seeking expert opinions.

Scientific Research and Teaching

Provides end-to-end reference implementation; open-source features facilitate reproduction and expansion.

Patient Education

Understand imaging results under doctor's guidance; local data retention ensures privacy.

Current Limitations

• Compromised model complexity: Browser resource constraints affect the recognition of rare/complex cases;
• Data format support: Need to enhance native support for medical standard formats like DICOM;
• Regulatory compliance: Need to pass medical device certifications like FDA/CE/NMPA.

Improvement Directions

• Model lightweighting: Knowledge distillation, pruning, and quantization to improve efficiency;
• Multimodal fusion: Integrate MRI sequences and other modalities like CT/PET;
• Federated learning integration: Aggregate multi-center data under privacy protection;
• 3D volume analysis: Expand to full 3D image analysis.

Architecture Insights

The privacy-first paradigm of "data stays, model moves" provides references for sensitive fields like healthcare and finance. WebAssembly/WebGPU technologies promote the expansion of client-side AI capabilities.

Conclusion

BrainVision AI demonstrates that privacy protection and model performance can coexist in medical AI. It provides a reference solution for relevant researchers and developers; open-source code lays the foundation for community exploration, and it is expected to play a role in more medical scenarios in the future.